← All field notes

Building an AI Assistant for Client Work? Keep Decisions Out of the Guesswork.

Design AI workflow automation for client briefs and proposals with source-linked facts, private records and human approval of scope, prices and dates.

Nov 11, 2026 · 5 min read · Updated Nov 14, 2026


Field Notes from Thunkle, a studio that takes AI-built apps from prototype to secure, production-ready software.

A client message can contain a price, a possible launch date and a feature request without any of them being agreed. AI workflow automation for client work needs to preserve that distinction before it prepares a proposal or meeting brief.

An assistant that turns those statements into a confident proposal has not saved useful work. It has made a commitment the person sending it may never have intended.

I am building Francolin around client context and proposal preparation. This article describes the design decisions I want that workflow to support, not a claim that every feature is shipped or that a measured time saving has been established. The same decisions matter when I scope custom AI workflows through Thunkle.

Define the useful output first

“Summarize this client” is a weak specification. Preparing for a particular call is much more concrete.

For a fictional website project, the accepted work covers public pages and a contact form. The client later asks about appointment booking and payment. Friday appears as a possible launch date, but no provider or additional price has been approved.

A useful brief distinguishes the current agreement, new requests, unresolved questions and decisions needed on the call. It should not describe bookings, payment and Friday as an accepted delivery.

Start with that output. Then decide which records, permissions and product states are necessary to produce it.

Store decisions separately from mentions

A retrieved sentence is evidence that something was said. It is not automatically evidence that it was approved.

Keep the source, speaker, date, project and decision state with consequential information. Useful states may include proposed, questioned, approved, superseded and unresolved. Choose states around the real workflow rather than building a large taxonomy nobody maintains.

An approval also needs an object. Agreement with a design is not acceptance of a price discussed elsewhere.

If two records conflict, show the conflict. Do not instruct the model to choose whichever value makes the document look complete.

Make retrieval reviewable

Source links should open the relevant record with enough context to check the claim. A link to the right thread can still support an incorrect summary if the assistant dropped a condition.

Review prices, delivery dates, scope and approval responsibility before use. Those statements can change what the business promises.

Keep a last-reviewed point and account for newer decisions. A brief based on last week's material should not silently appear current after yesterday's change.

For a support call, retrieve the failure and attempted fixes. For a returning-client proposal, retrieve the accepted scope, approved changes, delivered work and relevant commercial terms. The same client does not need the same summary every time.

Use history without copying the old deal

The previous invoice total does not explain the previous project.

A one-off discount, extra revision, delayed client input or included support can all change what that amount means. Retain those conditions beside the accepted scope.

In the fictional booking project, the public design might be reusable while accounts, payment and rescheduling need new estimates. The software should help identify that difference, not copy the old total onto a larger job.

Leave prices and dates blank when the operator has not supplied or approved them. A polished empty field is less dangerous than a plausible invented commitment.

Keep uncertain integrations uncertain

An unfamiliar booking provider may not support the required cancellation or payment behaviour.

The workflow should preserve that open question until somebody checks it. A generated proposal must not turn “investigate provider support” into “integration included”.

A paid discovery step can produce a recommendation and implementation estimate. That is a commercial decision for the person doing the work, not a number inferred from generic market rates.

Record dependencies on access, third-party approval and client-supplied material. The schedule should not begin on assumptions the client cannot see.

Separate preparation from sending

Drafting a reply, approving it and sending it are different actions.

For a scope change, the assistant can prepare the current agreement, proposed addition, questions, fee and delivery implications. The operator checks those facts and authorizes the actual offer.

Keep the approved version with its source decision. If the client agrees, update the delivery plan as well as the conversation history.

A positive comment about the idea does not accept a price that was never stated. Similarly, a call recap should not say “as agreed” while confirmation is pending.

Treat client context as private application data

Only process records the business is authorized to use. Restrict them by client, project and user role.

Test whether an unrelated account can retrieve material indirectly through search, summaries, generated documents or cached context. A secure list page is not enough if a retrieval route searches everybody's records.

Minimize personal data and secrets sent to external services. Decide retention, deletion and access revocation for source records and derived outputs.

The private-file access guide explains why summaries and processed outputs belong in the same authorization review as their source documents.

Test for invented commitments

Build synthetic examples with a proposed date, later conditional approval, a superseded price and an unresolved feature request.

Check whether the output preserves those distinctions and points to the right sources. Include a removed project member and an unrelated customer in access tests.

The acceptance criterion is not “the summary sounds good”. It is that consequential claims remain checkable, uncertain decisions remain open and private records stay with the correct audience.

A Thunkle AI workflow development engagement can scope this kind of retrieval, review and approval workflow in your existing product. The brief should name the decision the assistant supports, what it may prepare and what it must never commit on its own.

Build a client workflow that keeps decisions reviewable.

I can scope retrieval, integrations, permissions and approval steps around the work you already do, with clear limits on what the assistant may decide.